Intel's Cascade Lake is punctuated by 9200-series processors that come with a market-leading 56 cores and 112 threads. But those chips are destined for custom OEM systems only, leaving us to focus on the more accessible high-volume products.
Intel’s standard Cascade Lake models offer many of the iterative improvements that we expect from a generational refresh, such as a step up to faster memory data transfer rates, more memory capacity, higher clock rates, and slightly more performance-per-watt. However, they also lack the architectural refinements that would yield radical changes in performance. Intel does offer a few new features like support for new instructions that boost inference performance and hard-wired security mitigations to address some of the latest vulnerabilities. But the biggest advances involve other platform-level additives, along with Intel's pricing.
Intel bumped up Cascade Lake's clock rates, including the all-important Turbo Boost and AVX frequencies, across the breadth of its product stack. Then it left pricing alone, matching its previous-gen models. Those elevated clock rates build on Intel’s performance advantage in lightly-threaded workloads. Even the 28-core Platinum 8280 can maintain a surprisingly high 4.0 GHz Turbo Boost rate on two cores. That’s an attractive proposition for customers that pay high software licensing fees.
With a hard limit of 28 cores for its standard XCC silicon, the high-end Cascade Lake models still feature the same number of cores, albeit faster ones, for the same price. But Intel also beefed up several key entry-level and mainstream models with up to four more cores and more L3 cache, again leaving pricing untouched.
These are needed improvements in the face of AMD’s EPYC contenders, which cost much less than comparable Intel models while delivering ~85% as much performance in many types of workloads. We saw the EPYC 7601 carve out convincing wins in several parallel workloads that scale well, like C-Ray and NAMD. Then it demonstrated solid power consumption behaviors, besting the competition from Intel. AMD’s platform also has other notable advantages, such as less-strict product segmentation and up to 128 PCIe lanes on both single- and dual-socket servers. That healthy helping of PCIe connectivity is becoming more important for data storage, networking, and accelerators.
Intel’s focus on developing technologies like Optane DC Persistent Memory DIMMs might prove to be a critical advantage. For now, these modules only work in tandem with Intel memory controllers present on Cascade Lake processors. Naturally, the company will enable support in future Xeons. But it goes without saying that we won’t see this technology enabled on AMD platforms. That gives Intel an advantage for data center administrators who need the absolute highest memory density attainable, often resulting in the need for fewer servers. Aside from the Silver series, almost all of Intel's stack supports this memory tech. Intel does charge a premium for the Xeons that accommodate beefier 2.5 and 4TB allocations of Optane DIMMs, though that's worth a bit of extra cost for some customers.
Intel also has the advantage of being a long-term market leader, making it the less-risky choice for administrators. Granted, that advantage is receding as AMD continues to expand (not to mention preparing EPYC Rome processors). The company already has working 7nm silicon in its labs, and the 64C/128T processors are expected to arrive in mid-2019.
In all, Intel’s Platinum processors still offer the highest performance in the broadest range of workloads. But they certainly don’t take home a universal performance crown. As with all of Intel’s recommended customer pricing (RCP) for its data center chips, these prices have little to do with what the company’s largest customers pay after volume discounts. But given that the AMD EPYC 7601 offers nearly the same, or more, performance than the Xeon 8280 in some workloads, Intel’s pricing could use an even tighter haircut.
Overall, Intel delivers an impressive line of data center chips and supporting technologies, and socket compatibility with existing LGA 4367 servers is attractive to customers looking for a faster-than-normal upgrade cycle to reap the benefits of silicon-based mitigations. The majority of Intel's customers will migrate from much older platforms, and those customers will realize incredible performance gains over Haswell and Broadwell-era platforms.
Image Credits: Tom's Hardware
MORE: Best CPUs
MORE: All CPUs Content